The bounded test completed on 17 October 2026 UTC. All 15 outputs passed structural validation, with no missing responses or truncated finals. Actual usage confirms cache reuse on 13 requests. Semantic review and verified storage remain separate outcomes.
11 outputs accepted unchanged; 2 clear prompt-rule conflicts and 2 cases requiring adjudication remain held.
Structural validation checks output format and identity. It does not establish substantive correctness. Assistant review is not a human-gold accuracy estimate; DB readback verifies storage and provenance.
Of 104,791 input tokens, 87,594 were cache reads, 13,476 were cache writes and 3,721 were ordinary input. Thirteen requests read the shared prefix and two wrote it.
The previous 73-request diagnostic reported zero cache-read tokens. The new 15 demonstrate actual reuse, but different inputs, prompt, schema/cache setup and output cap mean this is not a controlled accuracy or per-article cost comparison.
| Component | Tokens | Actual cost (USD) |
|---|---|---|
| Ordinary input | 3,721 | $0.000186050 |
| Cache reads | 87,594 | $0.000437970 |
| Cache writes | 13,476 | $0.000842250 |
| Output, including reasoning | 11,618 | $0.002904500 |
| Total | 116,409 | $0.004370770 |
The 11,618 output tokens include 9,774 reasoning tokens and 1,844 visible or other output tokens. The actual cost is below the conservative reservation, which assumed upper-bound input and a full output cap. That difference is not a measured caching discount.
Two outputs have clear prompt-rule conflicts: an office was treated as an entity target, and an action span lacked a verb. Two other outputs require adjudication over unspecified government responsibility and the interpretation of discussion/cooperation. These adjudication cases are not asserted to be proven errors.
All 15 outputs and inputs were reviewed. No raw outputs were rewritten. The 11 accepted attempts were inserted and exactly read back; source documents were not modified. These are assistant judgments without a human gold reference.
Across the previous 73 outcomes and this new 15: 88 paid outcomes, 58 accepted attempts saved with exact DB readback, 17 clear prompt-rule conflicts, 10 adjudication cases and 3 unscorable length failures. All three unscorable outcomes belong to the earlier diagnostics; the new test had none.
Combined actual API cost for these diagnostics: $0.056833720. This is not total shared-account or all-history ledger spending. Different selected inputs and configurations prevent treating the before/after counts as a measured accuracy improvement.
This is one bounded diagnostic using the versioned prompt and shared-prefix cache setup. It does not establish sustained monthly cache rates, whole-corpus semantic quality or a precise monthly cost forecast. Existing holds and previous paid outcomes remain preserved.
The completed full-corpus input audit covers the frozen October 7 snapshot. The previous 73 diagnostic results remain a separate historical comparison.